Continual Wi‑Fi sensing must trade off plasticity—learning new rooms or users—and stability—preserving previously learned knowledge. However, nodes often cannot retain full channel state information (CSI) due to privacy and storage constraints.
WiMoT (Wi‑Fi Moment Transport) addresses this by storing, for each previously seen domain, the per‑class feature mean and covariance instead of raw samples. When the feature representation shifts, these stored statistics become stale. WiMoT limits the shift to an affine transformation on a frozen feature basis and applies the same transformation to the stored statistics, keeping them valid without needing old samples.
A supervised contrastive learning stage learns the new domain. A closed‑form loss on the stored statistics maintains separability of earlier classes. Afterwards a Gaussian discriminant classifier is refitted from the transported statistics, weighting each domain equally, so prediction does not require domain identity.
On the 3DO, WiMANS, and Widar3.0 benchmarks, WiMoT achieves average final accuracies of 94.02%, 89.61%, and 95.16% with forgetting at most 0.58 pp. Compared with 21 baselines (the strongest using replay), WiMoT attains the highest balance of new‑ and earlier‑domain accuracy on 3DO and Widar3.0, leading by 3.77 pp and 13.11 pp respectively, and remains ahead across domain orders and label budgets. On WiMANS it matches the strongest replay baseline (+0.41 pp) while reducing forgetting from 2.78 pp to 0.00 pp, and obtains the highest six‑order average accuracy.
Full covariances dominate memory usage; WiMoT stores them as rank‑32 factors, shrinking moment memory to 6.4% of the original size with at most 0.49 pp accuracy loss, fitting edge‑node storage budgets.
Review: WiMoT’s moment‑based storage enables replay‑free continual learning with strong performance‑memory trade‑offs, offering a practical solution for resource‑constrained Wi‑Fi sensing devices.